3-lead ECG combinations achieved equivalent AUROCs to 12-lead models for detecting LVSD (0.91), HCM (0.93), and CA (0.94) across external datasets.
Can AI models using combinations of up to 3 ECG leads achieve equivalent diagnostic performance to 12-lead models for detecting LVSD, HCM, and CA?
AI models using specific 3-lead ECG combinations can achieve diagnostic performance equivalent to 12-lead ECGs for detecting structural heart diseases, though the optimal leads vary by condition.
Absolute Event Rate: 0% vs 0%
Abstract Introduction Artificial intelligence (AI) models can accurately detect cardiovascular diseases such as left ventricular systolic dysfunction (LVSD)1,2, hypertrophic cardiomyopathy (HCM)3 and cardiac amyloidosis (CA)4 from 12-lead electrocardiogram (ECG). However, the optimal combination of leads remains unclear for wearable ECG device applications. Purpose To identify the optimal combinations of ECG leads for AI-based detection of LVSD, HCM and CA. Methods ECGs were collected from INST1 from US and INST2 from Japan. AI models to detect LVSD, HCM, and CA were trained at INST1 and externally validated at INST2. LVSD dataset was constructed using ECGs recorded within 14 days of an echocardiogram (echo). HCM and CA datasets were constructed by matching case ECGs to control ECGs in a 1:5 ratio based on age and sex. The datasets from INST1 were randomly divided into derivation, validation, and test sets in a 5:2:3 ratio without patient overlap. While all available ECGs in derivation and validation sets were used, only one ECG per patient was selected for the test set (the ECG closest to the echo for LVSD and the first available ECG after diagnosis for HCM and CA). To identify the optimal leads, all combinations of up to 3 leads were extracted from 12 leads and convolutional neural network models to detect LVSD, CA and HCM were trained with the same hyperparameters for 30 epochs. The final model for each lead combination was chosen as the model with the highest area under the receiver operating curve (AUROC) on the validation set. To account for variance by the initialization vector, the procedure was repeated 5 times for each lead combination, and 95% confidence intervals (CI) were calculated. Results A total of 75,033, 24,451, and 23,592 ECGs from INST1 and 11,084, 2,611, and 2,685 ECGs from INST2 were identified for LVSD, HCM and CA, respectively. The AUROCs for the12-lead models were 0.89 (95%CI, 0.90-0.88), 0.92 (0.90-0.93), and 0.94 (0.93-0.94) in the INST1 test sets and 0.91 (0.90-0.91), 0.93 (0.91-0.94), and 0.93 (0.91-0.94) in the INST2 external test set, respectively for LVSD, HCM and CA. Various 2- and 3-lead combinations showed equivalent AUROCs with the 12-lead model across diseases (Fig.1). The top 3-lead combination in each disease demonstrated consistent AUROCs in both INST1 and INST2: I-aVR-V6 for LVSD (INST1: 0.89 95%CI, 0.89-0.90; INST2: 0.91 0.91-0.92), I-II-V4 for HCM (INST1: 0.92 0.91-0.93; INST2: 0.93 0.93-0.93), and I-aVF-V2 for CA (INST1: 0.94 0.94-0.95; INST2: 0.94 0.92-0.94). The mean differences between AUROC of 12-lead models and that of each top-performing 3-lead model showed that the best combination of leads in one disease did not achieve 12-lead equivalence AUROCs for other diseases (Fig.2). All of the top ten 3-lead models included both limb and chest lead. Conclusion While 3-lead combinations achieved 12-lead equivalent AUROCs, the specific leads required may differ with the target disease.Fig.1 Fig.2
Nakayama et al. (Sat,) reported a other. 3-lead ECG combinations achieved equivalent AUROCs to 12-lead models for detecting LVSD (0.91), HCM (0.93), and CA (0.94) across external datasets.